Abstract
Alzheimer's disease (AD) is a progressive neurodegenerative disorder affecting approximately 50 million individuals worldwide, characterised by cognitive decline and memory impairment. Early and accurate staging is critical for timely intervention and treatment planning. This study proposes a novel hybrid deep learning architecture that combines a three-dimensional Convolutional Neural Network (3D CNN) with a Bidirectional Long Short-Term Memory (BiLSTM) network incorporating a self-attention mechanism for comprehensive five-stage dementia classification from magnetic resonance imaging (MRI) data. The model is trained and evaluated on combined ADNI and OASIS-2 datasets (3482 T1-weighted MRI scans) following a principled label-harmonisation procedure. A comprehensive preprocessing pipeline including skull stripping, bias field correction, spatial normalisation, and tissue segmentation is applied prior to modelling. The proposed CNN-LSTM model achieves 96.2% accuracy, 95.8% precision, 95.5% recall, and 95.6% F1-score across five cognitive stages: Cognitive Normal (CN), Significant Memory Concern (SMC), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and Alzheimer's Disease (AD). McNemar's test confirms statistically significant improvements over all baselines (p <0.01). Ablation studies, ROC analysis (macro AUC = 0.972), and attention weight visualisation demonstrate the contribution and interpretability of each architectural component. The model's potential for clinical deployment as a second-reader decision support tool is discussed, including inference time, GPU requirements, and planned validation strategies.